Ricursive's 65B Ripple Preview Scores 80.52 on Terminal-Bench 2.1 for a Claimed $32M, Edging a 501B Rival
Ricursive Intelligence released Ripple-65B Preview on November 4, a 65-billion-parameter dense model under Apache 2.0 that the company says sets a new US open-weight state of the art on agentic coding: 80.52 on Terminal-Bench 2.1.
The margin over the previous leader is thin. Reflection Beam, a 501B model, scores 80.1. The rest of the US open-weight field sits well back, by Ricursive’s own comparison table.
Key Numbers
| Model | Parameters | Terminal-Bench 2.1 |
|---|---|---|
| Ripple-65B Preview | 65B dense | 80.52 |
| Reflection Beam | 501B | 80.1 |
| Meta Muse Spark Open | 816B | 76.4 |
| Thinking Machines Inkling | 975B | 63.8 |
| NVIDIA Nemotron 3 Ultra | 550B | 56.4 |
- Pretraining: 12 trillion tokens, context up to 1M
- Reported all-in R&D cost: $32M, covering ladder experiments, data mixing, evals, synthetic data generation and the final pretraining and posttraining runs
- Post-training: 1.4M RL rollouts to reach 80.5, which Ricursive says is roughly 70x fewer than Reflection Beam needed
- Availability: Hugging Face, Together AI, GMI Cloud and OpenRouter at launch, with Fireworks following the same day
Why it is a dense model
Ricursive skipped mixture-of-experts. Ripple uses a thin-and-deep dense transformer (d_model 5120, 80 layers) with sliding-window and global attention layers, gated attention, global NoPE and value residuals. The company says it held about 40% MFU through pretraining and posttraining with simple sharding, no custom kernels and no quantization.
The pitch is local deployment. A 65B dense model has far fewer total weights than the 500B-plus MoE models it is compared against, and Ricursive names NVIDIA DGX Spark as a target device, a machine with plenty of compute but modest memory bandwidth.
The cost claim needs context
The $32M figure is Ricursive’s estimate and covers its full R&D spend. Its comparison point is a footnote: Reflection was reported to have committed $150M a month for compute at SpaceX’s Colossus cluster, and says Beam took four weeks to pretrain and four weeks to posttrain. Ricursive concludes its model cost about one-tenth of the prior open-weight leader. That is an inference from reported compute commitments, not an audited cost for Beam.
Ricursive also opted not to buy any post-training data, using its own synthetic pipelines and long-horizon synthetic environments. The RL recipe combines truncated importance sampling, score centering and a linear length penalty.
What to watch
The result is one benchmark, a preview checkpoint and a lab that is primarily a chip-design company; an internal variant of the model powers its design platform. A 0.4-point lead over Beam is inside the range where harness and scaffold choices can flip a ranking. Independent runs on Terminal-Bench 2.1, and results on benchmarks beyond it, will show whether a 65B dense model holds up against much larger open-weight MoE systems on real coding agents.